Open Models

Nex-N2.5 Mini Brings Open Weights to a New Agent Model Family

By Kaleido Field Staff · September 10, 2026

Citation-ready: Nex-N2.5 mini's official repository records a September 8 weight upload and Apache 2.0 metadata; the family's Max model is described as text-only, unlike mini and Pro.

Evidence boundary: Developer repository, files and self-reported capability descriptions. No local inference run, independent benchmark or guarantee of performance on a user's hardware.

What happened and why it matters

The model repository supplies weights and a reproducible revision anchor. That is stronger release evidence than a social announcement, but it does not independently validate the developer's benchmark comparisons.

The dated source

Primary reference: Nex-AGI model repository and dated commit history. Kaleido Field checked the event date and the article's attributed facts against this source.

The files, not the family name, define the release

The first-party API lists model weight shards and records repository creation on September 8. The retrieved revision is 87420286149d9cce9bd46cd335ef9bda33c37c1b. The model page identifies mini as 35B and labels it Apache 2.0.

The current card describes mini and Pro as multimodal, while Max uses a text-only foundation. A test requiring screen images therefore cannot assume that the largest member accepts the same inputs.

Runtime choices belong beside the result

The card documents adjustable reasoning effort and an example mini deployment using two H100 GPUs. Those are provider instructions, not a claim that this is the only possible hardware configuration or the cheapest deployment.

If testing a quantized or community-converted version, keep its provenance separate from the original weights. Record the template, tool parser and reasoning setting alongside the model revision so a failed action can be reproduced.

An agent score is not a completed user task

The benchmark tables are authored by the model developer. This article does not import those scores into an official leaderboard or rank consumer apps from them.

For an application pilot, use a bounded task and retain the tool calls, final artifact and acceptance check. The AgentCore evaluation report explains why a passing evaluation can still leave an untested permission boundary.

Evidence boundary

Developer repository, files and self-reported capability descriptions. No local inference run, independent benchmark or guarantee of performance on a user's hardware.